Reasoning Model

What Is Reasoning Model?

A reasoning model is a large language model trained to work through a problem step by step before committing to an answer, spending extra inference compute on internal deliberation. That thinking phase, often hidden or summarized, raises accuracy on math, code, and multi-step planning tasks at the cost of higher latency and token spend. The gains are measurable: DeepSeek-R1 scored 79.8% pass@1 on the AIME 2024 math competition, slightly ahead of OpenAI's o1-1217 [1].

Key Takeaways

• Reasoning models trade speed and cost for accuracy. The same question can cost several times more and take noticeably longer than on a standard model.
• The thinking happens as generated tokens, so you pay for reasoning output even when the provider hides or summarizes it.
• Most providers let you set a thinking budget or effort level, which is the main lever for balancing quality against latency in production.
• For agent work, reasoning models shine at planning and debugging, while routine edits and tool calls are usually better served by a faster non-reasoning model.

How It Works

A standard large language model produces its answer directly, one token at a time, with no separate deliberation phase. A reasoning model is trained, typically with reinforcement learning on problems that have verifiable answers, to first emit a long chain of intermediate steps: restating the problem, trying an approach, checking it, backtracking when something fails. Only after that internal pass does it write the final answer. The technique grew out of chain-of-thought prompting, but instead of asking the model to think step by step, the behavior is baked into training and happens whether you request it or not. The training signal is strong enough on its own that reinforcement learning lifted DeepSeek-R1-Zero's AIME 2024 score from an initial 15.6% to 71.0% pass@1, reaching 86.7% with majority voting [2].

From an engineer's seat, the practical differences are cost and latency. Thinking tokens are billed like output tokens, and a hard problem can burn thousands of them before the first visible word appears. APIs expose controls for this, usually a maximum thinking budget or a coarse effort setting, and agent frameworks increasingly route tasks by difficulty: a reasoning model plans the migration, a cheaper model executes the individual file edits. Reasoning output also occupies the context window, which matters in long agent sessions where every turn adds history.

Example

A team runs an AI coding agent against a flaky integration test that three engineers have failed to diagnose. On a fast non-reasoning model, the agent patches the assertion, the test goes green, and the underlying race condition survives. Rerun with a reasoning model at high effort, the agent spends about ninety seconds thinking, traces the failure to two workers sharing a temp directory, and proposes isolating the fixture per worker. The run costs roughly ten times more in tokens, and it is the first fix that actually holds. The team keeps the reasoning model for diagnosis and planning steps and routes mechanical edits to the cheap model.

What People Get Wrong

The common mistake is treating reasoning as a universal upgrade and pointing the strongest reasoning model at everything. For extraction, summarization, formatting, and simple tool calls, the extra deliberation adds latency and cost without improving results, and long thinking can even make a model overcomplicate a trivial task. Reasoning models earn their price on problems with real structure: multi-step logic, unfamiliar bugs, architectural tradeoffs. Route by task difficulty instead of defaulting to maximum brainpower.

Test-time compute

Test-time compute is the general idea that a model can get smarter at inference time by spending more computation on a query, rather than only by being trained longer or made bigger. Reasoning models are the mainstream implementation: their thinking tokens are test-time compute in action. Other forms include sampling several candidate answers and picking the best one, or having the model critique and revise its own draft. The engineering consequence is that intelligence became a runtime dial. You can pay more per request for better answers on demand, which turns model quality into a cost-optimization problem per task rather than a single procurement decision.

FAQ

What is a reasoning model versus a regular LLM? Both are large language models. A reasoning model has an extra trained behavior: it generates an internal chain of steps before answering, which improves reliability on hard problems. A regular model answers directly, faster and cheaper, which is fine for most everyday tasks.

Do reasoning models hallucinate less? They make fewer errors on problems they can check internally, like math and code, because they can catch their own mistakes mid-thought. They still hallucinate facts. A confident, well-reasoned argument built on an invented premise is a very real failure mode.

When should an agent use a reasoning model? Planning, debugging, reviewing tricky diffs, and anything where a wrong answer is expensive to unwind. Skip it for routine edits, retrieval, and formatting, where a standard model is faster and just as accurate.

Sources

1. DeepSeek. "DeepSeek-R1: 79.8% pass@1 on AIME 2024, slightly surpassing OpenAI o1-1217." https://arxiv.org/abs/2501.12948. Accessed August 2026.
2. DeepSeek. "DeepSeek-R1: reinforcement learning raised R1-Zero's AIME 2024 score from 15.6% to 71.0% pass@1, 86.7% with majority voting." https://arxiv.org/abs/2501.12948. Accessed August 2026.